End of Eruption: Using Machine Learning to Identify Transitions in Volcanic Time Series Data

The timing and processes that govern large-scale changes in volcanic systems, and their timing, are not yet fully understood. In recent years, the amount of data available for analysing patterns of volcanic behaviour has increased significantly. However, recognising the large-scale changes in observations of seismicity, geodesy and gas measurements remains difficult. Moreover, there is no systematic definition for the end of a volcanic eruption.

Machine learning algorithms can be used to determine the timing of significant changes in large and noisy datasets. Furthermore, machine learning approaches have been used, in the fields such as healthcare, finance and meteorology, to recognise non-linear patterns in data that traditional analyses failed to detect. These methods can also indicate the relative importance of different inputs to the models. Machine learning represents an important tool for detecting when the pattern of seismic activity critically changes, potentially signalling the onset or end of volcanic activity, and which features of the seismic activity (e.g. event rate, magnitude, frequency content) characterises those changes. By identifying patterns in volcanic seismicity, we aim to obtain an independent constraint on the underlying complex and non-linear processes that govern volcanic activity. We aim to quantitatively assess which parameters of seismic data are the most reliable in identifying changes in volcanic activity.

We apply classification techniques, including Gaussian process classifiers and support vector machines, to several persistent or periodically-active volcanic systems which display different patterns of persistent or variable eruptive behaviour. We classify eruptive and non-eruptive patterns of behaviour and compare these classifications to visual manifestations of volcanic activity, such as ash emission. Proof-of-concept studies on datasets from Telica (Nicaragua) and Nevado del Ruiz (Colombia) show that a machine-learning approach can successfully be applied to detect the beginning and end of eruption sequences; and our approach may offer a robust tool to help determine when an eruption has ended.